Skip to main content
ELI5: an agent can only touch what its workflow grants. tools says which named tools; tool_policy/sandbox_profile say how much power those tools get; mcp_server_config plugs in external tool servers.

Granting tools

required tools must be available or dispatch fails; optional tools attach when present. Common deeda tool names: fs.read, fs.write, git, probes.run, redaction.scan, merge.request_slot, beam.issue.update, canvas.workpad. Provider runtimes add their own built-ins (e.g. Claude Code’s bash/read/write/edit/glob/grep/web_search — see the Anthropic page).

Power profiles (the blast-radius dial)

tool_policy scopes the tool surface; sandbox_profile scopes filesystem/network. Omitting both preserves the default (edit ≈ workspace write, no network). read-only is the right default for review/research agents; network requires an explicit sandbox.network allowlist:

Attaching MCP servers

MCP servers attach per participant in multi-agent workflows:
deeda’s own MCP server (deeda-think) exposes the Think knowledge tools and attaches automatically when the run payload requests it. Remote MCP servers are partial-compatible on the Anthropic Messages path (connection contract proven; full remote round-trip not yet probed) — test small before depending on one.

When to use what

See also